Conditionalization: Adapting forward-chaining planners to partially observable environments

We provide a general way to take forward-chainingplanners for classical planning domains and condition-
alize them, i.e., adapt them to generate policies for
partially observable planning domains.
For domain-configurable planners such as SHOP2,
TLPlan, and TALplanner, our generalization tech-
nique preserves the ability to use domain knowledge to
achieve highly efficient planning. We demonstrate this
experimentally in two problem domains. The experi-
ments compare PKS and MBP (two existing planners
for partially observable planning) and CondSHOP2, a
version of the HTN planner SHOP2 created by apply-
ing our conditionalization method.
To our surprise, PKS and MBP could solve only the
simplest problems in our test domains. In contrast,
CondSHOP2 solved all of the test problems quite eas-
ily. This suggests that the ability to use domain knowl-
edge may be not just desirable but indeed essential for
solving large problems in partially observable domains.